Reinforcement-learning-based robust optimal control for underactuated USV with prescribed performance

Under realistic marine environment, it is rather challenging to precisely model an underactuated unmanned surface vehicle (USV), which makes optimal control methods relying on precise system models infeasible. This paper presents a solution for handling system uncertainties that solely relies on the input and output signals of the USV. To achieve this objective, a prescribed performance reinforcement learning (RL) control method is constructed to ensure that the USV achieves optimal control performance with guaranteed tracking precision. By using the state transformation design with prescribed performance and the manual positioning method, the initial motion dynamics of the underactuated USV is transformed into a standard integral cascade form, thereby transforming constrained tracking errors into unconstrained stable states, despite unknown dynamics, thereby enabling the control scheme combined with prescribed performance to be applied to solve the optimal control problem of the underactuated USV. Additionally, a robust term is added to compensate for the approximation errors of the neural network. Theoretical analysis ultimately verifies that the entire RL control scheme can ensure the prescribed tracking accuracy while achieving the optimal cost. Numerous simulation results demonstrate its remarkable performance.